Wavelet statistics selection for steganalysis using image noise

We present statistical techniques for steganalysis of natural images potentially subjected to data hiding. Our hypothesis is that image noises have certain statistics, while data hiding schemes will alter them and generate entirely different noises that can be tracked with high order statistics. We extract noises from images by wavelet denoising(WD) which outperforms Gaussian smoothing(GS) with respect to images with good quality. In this paper, wavelet packet decomposition followed with band selection is adopted to obtain statistics of image noises instead of wavelet decomposition. Though principal component analysis(PCA) is usually used for feature extraction, we implement analysis of variance(ANOVA) to select sub-band which leads to better results. The classifier between cover-images and stego-images is built using support vector machines(SVMs) which have good generalization performance.